This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH][AARCH64]Use shl for vec_shr_<mode> rtx pattern.
- From: Renlin Li <renlin dot li at arm dot com>
- To: Marcus Shawcroft <marcus dot shawcroft at gmail dot com>
- Cc: "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>, Alan Lawrence <Alan dot Lawrence at arm dot com>
- Date: Thu, 30 Apr 2015 12:55:57 +0100
- Subject: Re: [PATCH][AARCH64]Use shl for vec_shr_<mode> rtx pattern.
- Authentication-results: sourceware.org; auth=none
- References: <553FAA25 dot 7010001 at arm dot com> <CAFqB+PyEUzhL+-53Abku=U5Rs9_uCG5obkvsfAZ5n7KJOvftMg at mail dot gmail dot com>
Hi Marcus,
On 29/04/15 13:06, Marcus Shawcroft wrote:
I think there is another issue here, this change:
if (BYTES_BIG_ENDIAN)
- return "ushl %d0, %d1, %2";
+ return "shl %d0, %d1, %2";
else
return "ushr %d0, %d1, %2";
is in the context of:
(define_insn "vec_shr_<mode>"
[(set (match_operand:VD 0 "register_operand" "=w")
(lshiftrt:VD (match_operand:VD 1 "register_operand" "w")
(match_operand:SI 2 "immediate_operand" "i")))]
You are right. This pattern has ambiguity. I have updated the patch, and
represent vec_shr as an upspec. This will prevent other rtx patterns
implicitly matching this one.
The new patch is attached, is it Okay to commit?
Regards,
Renlin Li
gcc/ChangeLog:
2015-04-30 Renlin Li <renlin.li@arm.com>
* config/aarch64/aarch64-simd.md (vec_shr): Defined as an unspec.
* config/aarch64/iterators.md (unspec): Add UNSPEC_VEC_SHR.
gcc/testsuite/ChangeLog:
2015-04-30 Renlin Li <renlin.li@arm.com>
* gcc.target/aarch64/vect-reduc-or_1.c: New.
The RTL describes a right shift of the bits within each element in the
vector while the optab expects a right shift of the elements within
the vector?
/Marcus
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 0557570..6304eae6 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -783,12 +783,13 @@
;; For 64-bit modes we use ushl/r, as this does not require a SIMD zero.
(define_insn "vec_shr_<mode>"
[(set (match_operand:VD 0 "register_operand" "=w")
- (lshiftrt:VD (match_operand:VD 1 "register_operand" "w")
- (match_operand:SI 2 "immediate_operand" "i")))]
+ (unspec:VD [(match_operand:VD 1 "register_operand" "w")
+ (match_operand:SI 2 "immediate_operand" "i")]
+ UNSPEC_VEC_SHR))]
"TARGET_SIMD"
{
if (BYTES_BIG_ENDIAN)
- return "ushl %d0, %d1, %2";
+ return "shl %d0, %d1, %2";
else
return "ushr %d0, %d1, %2";
}
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 1fdff04..498358a 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -278,6 +278,7 @@
UNSPEC_PMULL ; Used in aarch64-simd.md.
UNSPEC_PMULL2 ; Used in aarch64-simd.md.
UNSPEC_REV_REGLIST ; Used in aarch64-simd.md.
+ UNSPEC_VEC_SHR ; Used in aarch64-simd.md.
])
;; -------------------------------------------------------------------
diff --git a/gcc/testsuite/gcc.target/aarch64/vect-reduc-or_1.c b/gcc/testsuite/gcc.target/aarch64/vect-reduc-or_1.c
new file mode 100644
index 0000000..f5d9460
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/vect-reduc-or_1.c
@@ -0,0 +1,34 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-vectorize -fdump-tree-vect-all" } */
+/* Write a reduction loop to be reduced using whole vector right shift. */
+
+extern void abort (void);
+
+unsigned char in[8] __attribute__((__aligned__(16)));
+
+int
+main (unsigned char argc, char **argv)
+{
+ unsigned char i = 0;
+ unsigned char sum = 1;
+
+ for (i = 0; i < 8; i++)
+ in[i] = (i + i + 1) & 0xfd;
+
+ /* Prevent constant propagation of the entire loop below. */
+ asm volatile ("" : : : "memory");
+
+ for (i = 0; i < 8; i++)
+ sum |= in[i];
+
+ if (sum != 13)
+ {
+ __builtin_printf("Failed %d\n", sum);
+ abort();
+ }
+
+ return 0;
+}
+
+/* { dg-final { scan-tree-dump "Reduce using vector shifts" "vect" } } */
+/* { dg-final { cleanup-tree-dump "vect" } } */